I Mapped Where Every AI Agent Actually Sits. Most People Pick Wrong.
TL;DR
OpenClaw matters less as a product than as the map everyone else now has to play on — Nate’s core claim is that Perplexity, Meta’s Manis, Anthropic Dispatch, and even Lovable are all reacting to the category OpenClaw defined, not just copying features.
The real framework is three axes: where the agent runs, who orchestrates the intelligence, and what interface you use — local vs. cloud vs. hybrid shapes privacy and security, model routing shapes cost and lock-in, and messaging surface shapes whether the product actually fits your habits.
OpenClaw is the sovereignty bet: maximum control, maximum flexibility, and maximum risk — it runs locally, lets users plug in their own LLMs and messaging platforms, has roughly 250,000 users, but also faces serious security concerns like 30,000 exposed instances and 800 compromised skills from a supply-chain attack.
Perplexity Computer turns OpenClaw’s biggest weakness into its whole pitch: delegation — for $200/month, it offers cloud-based outcome-level work where Perplexity handles orchestration and security, making it attractive for knowledge workers and enterprises that want results without managing infrastructure.
Meta’s Manis is not an enterprise play so much as a distribution play for keeping attention inside the Meta ecosystem — after Meta’s $2 billion acquisition, Nate argues the point is to capture the “agent moment” for Zuck’s 3 billion users, even if privacy-sensitive users will balk.
Nate’s 2026 thesis is ‘relentless simplification’: agents are compressing the interface layer across software — tools will either go deep with unique capability or go broad as default delegation layers, while the middle ground is “where you go to die.”
The Breakdown
OpenClaw Is Bigger Than the Horse Race
Nate opens by arguing that most coverage has missed the point. People are obsessing over whether OpenClaw is beating rivals or whether it’s a “dumpster fire” of security issues, but the real story is that every major company has made a different strategic bet in response.
The Flood of Forks, Clones, and Corporate Responses
He rattles off just how crowded this has become: Nvidia’s Nemo Claw, OpenAI hiring Peter for an upcoming launch, Meta spending $2 billion on Manis, and Lovable pivoting toward the same category. On the open-source side, ZeroClaw rewrote it in Rust, Moltus targeted enterprise Rust deployments, Open Fang pitched an agent OS, and Nanobot stripped it to 4,000 lines of code — the classic Linux/Android dynamic where every weakness becomes a startup thesis.
The Three Axes That Actually Matter
Instead of the usual “more control vs. less control” framing, Nate says there are three questions that matter: where the agent runs, who orchestrates the intelligence, and what interface contract it assumes. That’s his practical filter for judging whether any new agent product is actually useful to you or just another confusing launch.
OpenClaw as the Sovereignty Play
OpenClaw runs locally, uses your API keys and data, and lets you mix your own LLMs and messaging platforms like Telegram, WhatsApp, Signal, and Slack. Nate frames it as a direct attempt to “disintermediate” platforms and put users in charge — but he pairs that with ugly security reality: over 30,000 exposed instances with weak or missing auth, plus a skills registry supply-chain attack with 800 compromised skills.
Perplexity Computer: Delegation for People Who Don’t Want the Headache
Perplexity takes OpenClaw’s complexity and flips it into a product: describe an outcome, let cloud agents handle the decomposition, and trust Perplexity to run it all in a secure virtual box. Nate jokes about the company’s positioning twists — including its move toward a more local “personal computer” option — but says the product is genuinely worth considering for knowledge workers and enterprise teams, especially if $200/month is less painful than managing your own infrastructure.
Manis After Meta: The Distribution Bet
Nate reads Meta’s move through Zuck’s favorite unit of analysis: eyeballs and hours in the day. His take is that Manis exists to capture agent usage inside Meta’s ecosystem for consumers and small businesses, not to win on privacy or model flexibility; the key question becomes whether you trust Meta with your data.
Anthropic Dispatch: Safety Branded as Product Design
Anthropic’s Dispatch is a simpler play: message Claude from your phone and drive Claude on your computer, with a secure, single-threaded setup. Nate calls it “primitive” compared with multimodel orchestration, but also incredibly practical for non-technical professionals who already trust Claude and just want to browse, make a PowerPoint, run email, or manage a calendar without setting up an OpenClaw stack.
Lovable, Agent Compression, and the Real Question of 2026
The most ironic case, he says, is Lovable — the vibe-coding darling past $300 million ARR now being forced to evolve into a more general execution tool because agents are swallowing software categories. That leads to his big thesis: agents are “compressing the interface layer,” so the winners either go deep with unique capability or go broad as the default delegation layer, while the middle gets wiped out. His closing lens is blunt and memorable: the big question of 2026 is how we choose to delegate agentic trust.